The Pseudosquares Sieve will drop the memory requirements much further from sqrt(n) to log^2(n). https://link.springer.com/chapter/10.1007/11792086_15

Yep, this is the natural way to go, especially considering the possibility of parallel computing and the importance of cache locality, etc.